nodejs benchmark running only on one core about benchmarks

[kazu-yamamoto]

Just FYI.

When this benchmark was created, this feature of nodejs was not available.

And the current GHC does not make use of cores well even if -N is specified since there is just one IO manager. Now, the parallel IO manager is merged to GHC head. So, if we use GHC head, Yesod scales on multi cores well.

[benma]

Thanks.

Still, then node should have been launched 3 times on three different ports and having httperf fire at all of them (or the Haskell versions should have been run with -N1), to ensure a comparison on the same grounds.

On GHC 7.4.2, I see a two-fold increase in performance when going from -N1 to -N3. Also, I see multiple cores maxing out in htop. Is the old IO manager the reason that I don't see a three-fold increase?

[kazu-yamamoto]

GHC 7.4.2 is using the old manager. Yes, we cannot expect that -N3 results in a three-fold increase.

[kazu-yamamoto]

You may be interested in this article: https://github.com/snoyberg/posa-chapter/blob/master/warp.md

[benma]

Thank you. So mightyhttpd2 uses forkIO to run on multiple cores. Sorry for the noob question, but how would I adapt the warp.hs benchmark script to also fully utilize all cores (without the new IO manager)?

[kazu-yamamoto]

Mighttpd2 uses forkProcess to run on multiple cores. I'm not the author of this package. Decision should go to @snoyberg.

[snoyberg]

I have no problem merging in this patch, the point of the benchmarks was simply to use the tools we were comparing against in their standard configuration (thus the reason that Yesod doesn't use the prefork technique). If you send a pull request, I'll merge it.

[benma]

I sent the pull request. Could you please tell me how to modify the warp example to use the prefork technique as well? I would like to see how it fares until the parallel IO manager is released.

[snoyberg]

You can see some code I put together at:

https://gist.github.com/snoyberg/5335726

But @kazu-yamamoto is really the pre-fork expert; I've been meaning to ask him to have a look at that file and see if it makes sense.

[kazu-yamamoto]

If this runs as a foreground process (not as a daemon), this code looks good. Please don't forget to remove the +RTS -N option.

[benma]

Thanks a lot to both of you. I appreciate the help!